航拍多光谱田间秸秆覆盖量反演模型的建立与优化  

Establishment and optimization of aerial multispectral field straw mulch quantity inversion model

作　　者：刘媛媛[1] 孙宇 高雪冰 王利斌 王跃勇 刘梦琪 崔舒然 LIU Yuanyuan;SUN Yu;GAO Xuebing;WANG Libin;WANG Yueyong;LIU Mengqi;CUI Shuran(College of Information Technology(Institute of Intelligent Agriculture),Jilin Agricultural University,Changchun 130118,China;College of Engineering and Technology,Jilin Agricultural University,Changchun 130118,China;Changchun Agricultural Machinery Research Institute,Changchun 130052,China;Jilin Agricultural Machinery Research Institute,Changchun 130022,China)

机构地区：[1]吉林农业大学信息技术学院(智慧农业研究院),吉林长春130118 [2]吉林农业大学工程技术学院,吉林长春130118 [3]长春市农业机械研究院,吉林长春130052 [4]吉林省农业机械研究院,吉林长春130022

出　　处：《光学精密工程》2024年第11期1773-1787,共15页Optics and Precision Engineering

基　　金：国家自然科学基金资助项目(No.42001256);吉林省科技厅技术创新引导项目(No.20220402023GH);吉林省科技厅重点研发项目(No.20230202039NC)。

摘　　要：保护性耕作是农业耕地可持续性发展的重要方法,已被世界多地采用,秸秆覆盖量实现从“有无”到“多少”的进一步判定,是秸秆还田检测的重要指标。通过无人机搭载多光谱相机航拍研究区内春秋两季遥感数据,并同步测定玉米秸秆覆盖量。首先,通过遥感数据提取光谱反射率并构建光谱指数,采用相关系数法筛选出对秸秆覆盖量敏感的波段变量和光谱变量,作为模型输入变量;然后,采用支持向量机(Support Vector Machine,SVM)、随机森林(Random Forest,RF)、BP神经网络(Back Propagation Neural Network,BPNN)和极限学习机(Extreme Learning Machine,ELM)4种机器学习算法,建立玉米秸秆覆盖量的反演模型,比较不同时期和不同研究区域的模型精度;最后,为解决预测性能受其模型参数影响较大问题,引入遗传算法(Genetic Algorithm,GA)和粒子群算法(Particle Swarm Optimization,PSO),并提出遗传-粒子群混合算法(Genetic-Particle Swarm Optimization,GA-PSO),利用它们的互补性提高模型的性能,完成区域内秸秆覆盖量的估算。实验结果表明,基于GA-PSO优化的RF算法玉米秸秆覆盖量反演模型取得了最佳的反演效果,其中R^(2)达到了0.74。同时,对比分析不同数据的反演结果,均较为真实地反映了区域内秸秆覆盖量,估测准确率达到91.36%,说明可以通过优化模型实现结果估算。研究为保护性耕作秸秆还田量检测提供科学参考,亦为其他作物秸秆覆盖量估测提供了可靠的模型反演方法。Conservation tillage is a crucial method for the sustainable development of agricultural arable land and has been adopted worldwide.The quantity of straw mulch is determined not just by its presence but by its amount,serving as a key indicator for detecting straw return to the field.In this study,aerial remote sensing data from the spring and autumn seasons were captured using a UAV equipped with a multi-spectral camera,while the corn straw mulch quantity was measured simultaneously.Spectral reflectance was first extracted,and spectral indices were constructed from the remote sensing data.The correlation coefficient method was then used to identify the band variables and spectral variables sensitive to the straw mulch quantity,which served as model input variables.Subsequently,machine learning algorithms such as support vector machine(SVM),random forest(RF),BP neural network(BPNN),and extreme learning machine(ELM)were employed to establish the inversion model for straw mulch quantity.The accuracy of these models was compared across different time periods and study areas.To address the significant impact of model parameters on predictive performance,genetic algorithm(GA)and particle swarm optimization(PSO)were introduced,culminating in the proposed genetic-particle swarm optimization hybrid algorithm(GA-PSO).This hybrid approach leveraged their complementary strengths to enhance model performance and complete the estimation of straw coverage in the region.The results indicated that the RF algorithm optimized by GA-PSO achieved the best inversion effect for corn straw mulch quantity,with an R^(2)value of 0.74.Comparative analysis of different data sets consistently reflected the straw mulch quantity in the region accurately.The accuracy of estimating the corn straw mulch quantity in the field reached 91.36%,demonstrating that result estimation can be effectively achieved through model optimization.This study provides a scientific reference for detecting straw return in conservation tillage and offers a reliable mod

关 键 词：多光谱图像 机器学习 秸秆覆盖量 无人机 遗传算法 粒子群算法 

分 类 号：S24[农业科学—农业电气化与自动化] TP79[农业科学—农业工程]